The Rhetoric of Atrocity: Modeling Genocidal Speech

Abstract

This dissertation examines genocidal hate speech in the Tigray conflict through a multi-method lens, combining the development of a supervised classification model with human affective validation. In the first study, Tweets were harvested using a domain-specific lexicon, capturing messages containing ethnic slurs and conflict-related keywords. Unsupervised topic modeling and lexical feature analysis revealed distinct narratives, and building on these insights, a HateBERT classifier was fine-tuned to assign tweets to a ten-category taxonomy rooted in Atrocity Speech Law theory (Gordon (2017)). In a second study, members of the Tigrayan diaspora rated a subset of tweets on anger, fear, disgust, perceived threat, perceived violence and behavioral intentions. Repeated-measures analyses showed that direct calls for violence and dehumanizing metaphors elicited the strongest negative reactions, whereas economic insults provoked relatively mild responses. The convergence between model outputs and human judgments supports the taxonomy’s validity. The findings underscore the potential of theory-informed NLP models for early warning and moderation, while highlighting the need for richer context, multilingual capabilities and continuous human feedback.